978-1-4673-9098-9/15/$31.00 ©2015 IEEE 293
2015 8th International Congress on Image and Signal Processing (CISP 2015)
Nonlocal Means Denoising Using a Content-Based
Searching Region
Guangyu Xu, Shexiang Jiang
School of Computer Science and Engineering
Anhui University of Science and Technology
Huainan, 232001, China
Abstract—The recent Nonlocal Means (NLM) filter shows
excellent performance for image denoising. We investigate the
effect of two key parameters on NLM: the searching region size
and the decay parameter of the weights. We testify the general
knowledge that the denoising performance can be improved by
adjusting the searching region size and using an optimum decay
parameter. Based on this observation, we adaptively choose the
size of the searching region according to the edge contents of an
area, and present a linear estimation for the decay parameter.
Experimental results show that the proposed algorithm achieves
much better denoising performance compared with the original
NLM filter.
Keywords-image denoising; NLM; adaptive searching window;
edge contents
I. INTRODUCTION
The purpose of image denoising is to restore the high-quality
clean image from its noisy version. As an important research
area in the image processing community, it has attracted
researcher’s attention for many years and is still an unsolved
problem till now. This is because the distinguishing the noise
and the image details is very difficult. The image details are
often misjudged to the noise and over-smoothed or even
removed in the process of the restoration.
Buades et al. [1] compared several classic filtering methods
and proposed the nonlocal means (NLM) algorithm for image
denoising. They show that the NLM filter outperforms some
other classic filtering methods, such as anisotropic diffusion [2],
total variation [3], bilateral filtering [4], and wavelet-based
denoising [5]. The NLM filter exploits spatial redundancy or
self-similarity in an image for noise removal and can give rise
to promising results. For each pixel, the estimate value is
achieved by the weighted average of the intensities of all the
pixels within the image. The weights are proportional to the
similarities of the patch centered at the target pixels and the
patches centered at other pixels. Generally, better noise
removal can be obtained if a pixel can find much more similar
candidates for the weighted average.
Many researches based on NLM (patch-based) method have
been presented recently. In [6], Kervrann et al. improved
NLM by automatically and locally selecting the local
searching region size. The optimal size is found by iteratively
grow the local searching region size and terminate the iteration
according to some rules. In [7], the weight function, which is
based on probabilistic block, updates the weights in a data-
driven approach using both noisy blocks and iteratively
filtered results. In order to obtain similar patches, the rotated
patch matching strategy in [8] was employed. However, it
only used four angles for patch rotation. Mahmoudi and
Sapiro [9] proposed an approach for speeding up the NLM
filtering with filters that discard unrelated neighbors from the
weighted average. In [10], an improved NLM method that
achieves a higher denoising performance and lower
computational load is proposed by using principal component
analysis (PCA). In order to optimize the NLM parameters,
Dimitri and Michel [11] used the Stein’s unbiased risk
estimate (SURE) to control the mean square error (MSE) for
image filtering and thus selected the optimal decay parameter
for the weight function. Salmon [12] discussed the choices of
another two parameters: the searching window size and the
weights gave to the pixels of the central patch, and gave some
valuable suggestions.
Azzabou et al. [13] suggested partitioning the image into
two classes: noisy smooth regions and noisy edge/texture
regions. Each pixel is characterized with a statistical model
that defines a membership degree to each class. Sun et al. [14]
presented a method that determines a pixel-wise adaptively-
shaped searching area, within which images are mainly
homogeneous. This method can effectively eliminate the
dissimilar candidates in the searching region from the
weighted average.
In this paper, we propose an improved NLM method by
adaptively choosing the searching region size for image
denoising. We study the influence of the searching region size
in the original NLM algorithm. For the edge or texture region,
a large searching region size can preserve more image details.
Conversely, a searching region with small size for the smooth
region can remove noise effectively. The optimal size of local
searching region is determined through taking into account the
edge contents of an area. In addition, we also present a linear
estimation for the decay parameter to obtain better filtering
results. Here, we do not aim to decrease the computational
time, and only are concerned about the denoising performance
of NLM filter.
The rest of this paper is organized as follows. In Section II,
the NLM method is briefly reviewed. Section III describes the